AK Ashutosh Kakadiya
CV
Research

Research

My work sits at the intersection of reinforcement learning, statistical relational learning, and deep generative models — with a bias toward interpretable methods and real-world impact.

01

Relational Contextual Bandits in real-world user interactions

Contextual bandit algorithms have become essential in real-world user-interaction problems, but they represent context as attribute–value pairs, making them infeasible for inherently relational domains like social networks. We propose Relational Boosted Bandits (RB2), a contextual bandits algorithm for relational domains based on relational boosted trees. RB2 learns interpretable and explainable models thanks to the descriptive nature of relational representation, and is effective on link prediction, relational classification, and recommendation.

AAAI’21 paper
02

Correlated discrete data generation using adversarial training

GANs have shown great promise for synthetic image generation, inpainting, and style transfer, but generating discrete data remains a challenge. This work presents an adversarial-training-based correlated discrete data (CDD) generation model, with a detailed approach for conditional CDD generation. Evaluated on a job-seeking candidate skill-set dataset and MNIST, the model outperforms prior work by leveraging the inherent correlation in the data.

Project page
03

Human Capital Management recommendation system

Finding meaningful work and the right talent is a classic Human Capital Management problem. This project builds a stateless, scalable, micro-service architecture for automated HCM, using ML to cluster and categorize job postings and candidate profiles. NLP handles feature extraction while a semantic-web ranking algorithm produces accurate recommendations.